Exploratory Analysis for Frontal Lobe
#setwd('/afs/inf.ed.ac.uk/user/s17/s1725186/Documents/PhD-Models/FirstPUModel/RMarkdowns')
library(tidyverse) ; library(reshape2) ; library(glue) ; library(plotly) ; library(plotlyutils)
library(RColorBrewer) ; library(viridis) ; require(gridExtra) ; library(GGally)
library(Rtsne)
library(ClusterR)
library(DESeq2) ; library(biomaRt)
library(knitr)Load preprocessed dataset (preprocessing code in 20_02_21_data_preprocessing.Rmd)
# Gandal dataset
load('./../Data/preprocessed_data.RData')
datExpr = datExpr %>% data.frame
DE_info = DE_info %>% data.frame
# GO Neuronal annotations: regex 'neuron' in GO functional annotations and label the genes that make a match as neuronal
GO_annotations = read.csv('./../Data/genes_GO_annotations.csv')
GO_neuronal = GO_annotations %>% filter(grepl('neuron', go_term)) %>%
mutate('ID'=as.character(ensembl_gene_id)) %>%
dplyr::select(-ensembl_gene_id) %>% distinct(ID) %>%
mutate('Neuronal'=1)
# Update DE_info with Neuronal information
DE_info = DE_info %>% mutate('ID'=rownames(.)) %>% left_join(GO_neuronal, by='ID') %>%
mutate(Neuronal=ifelse(is.na(Neuronal), 0, Neuronal)) %>%
mutate(significant=padj<0.05 & !is.na(padj))
rm(GO_annotations)All samples together
- There seem to be two different behaviours in mean expression by gene
plot_data = data.frame('ID'=rownames(datExpr), 'Mean'=rowMeans(datExpr))
p1 = ggplotly(plot_data %>% ggplot(aes(Mean)) + geom_density(color='#0099cc', fill='#0099cc', alpha=0.3) +
scale_x_log10() + theme_minimal())
plot_data = data.frame('ID'=colnames(datExpr), 'Mean'=colMeans(datExpr))
p2 = ggplotly(plot_data %>% ggplot(aes(Mean)) + geom_density(color='#0099cc', fill='#0099cc', alpha=0.3) +
theme_minimal() + ggtitle('Mean expression density by Gene (left) and by Sample (right)'))
subplot(p1, p2, nrows=1)rm(p1, p2, plot_data)Grouping genes by Neuronal tag and samples by Diagnosis
The two groups of genes seem to be partially characterised by genes with Neuronal function
In general, the autism group has a bigger mean and is more spread out than the control group
plot_data = data.frame('ID'=rownames(datExpr), 'Mean'=rowMeans(datExpr)) %>%
left_join(GO_neuronal, by='ID') %>% mutate('Neuronal'=ifelse(is.na(Neuronal),F,T))
p1 = plot_data %>% ggplot(aes(Mean, color=Neuronal, fill=Neuronal)) + geom_density(alpha=0.3) +
scale_x_log10() + theme_minimal() + theme(legend.position='bottom') +
ggtitle('Mean expression density by gene')
plot_data = data.frame('ID'=colnames(datExpr), 'Mean'=colMeans(datExpr)) %>%
mutate('ID'=substring(ID,2)) %>% left_join(datMeta, by=c('ID'='Dissected_Sample_ID'))
p2 = plot_data %>% ggplot(aes(Mean, color=Diagnosis, fill=Diagnosis)) + geom_density(alpha=0.3) +
theme_minimal() + theme(legend.position='bottom') +
ggtitle('Mean expression density by Sample')
grid.arrange(p1, p2, nrow=1)
rm(GO_annotations, plot_data, p1, p2)ASD vs CTL
In general there doesn’t seem to be a lot of variance in mean expression between autism and control samples by gene.
plot_data = data.frame('ID'=rownames(datExpr),
'ASD'=rowMeans(datExpr[,datMeta$Diagnosis=='ASD']),
'CTL'=rowMeans(datExpr[,datMeta$Diagnosis!='ASD']))
plot_data %>% ggplot(aes(ASD,CTL)) + geom_point(alpha=0.1, color='#0099cc') +
geom_abline(color='gray') + ggtitle('Mean expression ASD vs CTL') + theme_minimal()
Grouping genes and samples by Diagnosis
There doesn’t seem to be a noticeable difference between mean expression by gene between Diagnosis groups
Samples with autism tend to have a wider dispersion of mean expression with higher values than the control group (as we had already seen above)
plot_data = rbind(data.frame('Mean'=rowMeans(datExpr[,datMeta$Diagnosis=='ASD']), 'Diagnosis'='ASD'),
data.frame('Mean'=rowMeans(datExpr[,datMeta$Diagnosis!='ASD']), 'Diagnosis'='CTL')) %>%
mutate('Diagnosis'=factor(Diagnosis, levels=c('CTL','ASD')))
p1 = ggplotly(plot_data %>% ggplot(aes(Mean, color=Diagnosis, fill=Diagnosis)) +
geom_density(alpha=0.3) + scale_x_log10() + theme_minimal())
plot_data = rbind(data.frame('Mean'=colMeans(datExpr[,datMeta$Diagnosis=='ASD']), 'Diagnosis'='ASD'),
data.frame('Mean'=colMeans(datExpr[,datMeta$Diagnosis!='ASD']), 'Diagnosis'='CTL')) %>%
mutate('Diagnosis'=factor(Diagnosis, levels=c('CTL','ASD')))
p2 = ggplotly(plot_data %>% ggplot(aes(Mean, color=Diagnosis, fill=Diagnosis)) +
geom_density(alpha=0.3) + theme_minimal() +
ggtitle('Mean expression by Gene (left) and by Sample (right) grouped by Diagnosis'))
subplot(p1, p2, nrows=1)rm(p1, p2, plot_data)Visualisations
Samples
PCA
The first principal component seems to separate relatively well the two diagnosis, although not as good as when it was with all the brain regions
pca = datExpr %>% t %>% prcomp
plot_data = data.frame('ID'=colnames(datExpr), 'PC1' = pca$x[,1], 'PC2' = pca$x[,2]) %>%
mutate('ID'=substring(ID,2)) %>% left_join(datMeta, by=c('ID'='Dissected_Sample_ID')) %>%
dplyr::select('ID','PC1','PC2','Diagnosis') %>%
mutate('Diagnosis'=factor(Diagnosis, levels=c('CTL','ASD')))
plot_data %>% ggplot(aes(PC1, PC2, color=Diagnosis)) + geom_point() + theme_minimal() + ggtitle('PCA') +
xlab(paste0('PC1 (',round(100*summary(pca)$importance[2,1],1),'%)')) +
ylab(paste0('PC2 (',round(100*summary(pca)$importance[2,2],1),'%)'))
rm(pca, plot_data)MDS
Looks exactly the same as the PCA visualisation, just inverting the x axis
d = datExpr %>% t %>% dist
fit = cmdscale(d, k=2)
plot_data = data.frame('ID'=colnames(datExpr), 'C1'=fit[,1], 'C2'=fit[,2]) %>%
mutate('ID'=substring(ID,2)) %>% left_join(datMeta, by=c('ID'='Dissected_Sample_ID')) %>%
dplyr::select('C1','C2','Diagnosis') %>%
mutate('Diagnosis'=factor(Diagnosis, levels=c('CTL','ASD')))
plot_data %>% ggplot(aes(C1, C2, color=Diagnosis)) + geom_point() + theme_minimal() + ggtitle('MDS')
rm(d, fit, plot_data)t-SNE
Higher perplexities seem to capture the difference between diagnosis better. You can separate the diagonsis perfectly with a line using the highest perplexity
#perplexities = c(2,5,10,25)
perplexities = c(1,2,3,4)
ps = list()
for(i in 1:length(perplexities)){
set.seed(123)
tsne = datExpr %>% t %>% Rtsne(perplexity=perplexities[i])
plot_data = data.frame('ID'=colnames(datExpr), 'C1'=tsne$Y[,1], 'C2'=tsne$Y[,2]) %>%
mutate('ID'=substring(ID,2)) %>% left_join(datMeta, by=c('ID'='Dissected_Sample_ID')) %>%
dplyr::select('C1','C2','Diagnosis','Subject_ID') %>%
mutate('Diagnosis'=factor(Diagnosis, levels=c('CTL','ASD')))
ps[[i]] = plot_data %>% ggplot(aes(C1, C2, color=Diagnosis)) + geom_point() + theme_minimal() +
ggtitle(paste0('Perplexity=',perplexities[i])) + theme(legend.position='none')
}
grid.arrange(grobs=ps, nrow=2)
rm(ps, perplexities, tsne, i)Genes
PCA
First Principal Component explains over 99% of the total variance
There’s a really strong correlation between the mean expression of a gene and the 1st principal component
pca = datExpr %>% prcomp
plot_data = data.frame( 'PC1' = pca$x[,1], 'PC2' = pca$x[,2], 'MeanExpr'=rowMeans(datExpr))
plot_data %>% ggplot(aes(PC1, PC2, color=MeanExpr)) + geom_point(alpha=0.3) + theme_minimal() +
scale_color_viridis() + ggtitle('PCA') +
xlab(paste0('PC1 (',round(100*summary(pca)$importance[2,1],1),'%)')) +
ylab(paste0('PC2 (',round(100*summary(pca)$importance[2,2],1),'%)'))
rm(pca, plot_data)t-SNE
Higher perplexities capture a cleaner visualisation of the data ordered by mean expression, in a similar (although not as linear) way to PCA
perplexities = c(1,2,5,10,50,100)
ps = list()
for(i in 1:length(perplexities)){
tsne = read.csv(paste0('./../Visualisations/tsne_perplexity_',perplexities[i],'.csv'))
plot_data = data.frame('C1'=tsne[,1], 'C2'=tsne[,2], 'MeanExpr'=rowMeans(datExpr))
ps[[i]] = plot_data %>% ggplot(aes(C1, C2, color=MeanExpr)) + geom_point(alpha=0.5) + theme_minimal() +
scale_color_viridis() + ggtitle(paste0('Perplexity = ',perplexities[i])) + theme(legend.position='none')
}
grid.arrange(grobs=ps, nrow=2)
rm(perplexities, ps, tsne, i)Differential Expression Analysis
- Only 24 genes (~0.15%) are significant using a threshold of 0.05 for the adjusted p-value and a without a log Fold Change threshold (keeping the null hypothesis \(H_0: lfc=0\))
cat(paste0(sum(is.na(DE_info$padj)), ' (~', round(100*(sum(is.na(DE_info$padj))/nrow(DE_info))),
'%) of the p-values couldn\'t be calculated'))## 1880 (~12%) of the p-values couldn't be calculatedtable(DE_info$padj<0.05, useNA='ifany')##
## FALSE TRUE <NA>
## 14250 24 1880p = DE_info %>% ggplot(aes(log2FoldChange, padj, color=significant)) + geom_point(alpha=0.2) +
scale_y_sqrt() + xlab('log2 Fold Change') + ylab('Adjusted p-value') + theme_minimal()
ggExtra::ggMarginal(p, type = 'density', color='gray', fill='gray', size=10)## Warning: Removed 1880 rows containing missing values (geom_point).
rm(p)There is a clear negative relation between lfc and mean expression, for both differentially expressed and not differentially expressed groups of genes
The relation is strongest for genes with low levels of expression
The p-value of the genes with the lowest levels of expression couldn’t be calculated. DESeq2 does this with genes where it considers the level of expression is not high enough to get robust results
plot_data = data.frame('ID'=rownames(datExpr), 'meanExpr'=rowMeans(datExpr)) %>% left_join(DE_info, by='ID') %>%
mutate('statisticallySignificant' = ifelse(is.na(padj),NA, ifelse(padj<0.05, TRUE, FALSE)),
'alpha' = ifelse(padj>0.05 | is.na(padj), 0.1, 0.5))
plot_data %>% ggplot(aes(meanExpr, abs(log2FoldChange), color=statisticallySignificant)) +
geom_point(alpha = plot_data$alpha) + geom_smooth(method='lm') +
theme_minimal() + scale_y_sqrt() + theme(legend.position = 'bottom') +
xlab('Mean Expression') + ylab('abs(lfc)') + ggtitle('Log fold change by level of expression')
When filtering for differential expression, the points separate into two clouds depending on whether they are over or underexpressed
The top cloud corresponds to the over expressed genes and the bottom to the under expressed ones
datExpr_DE = datExpr[DE_info$significant,]
pca = datExpr_DE %>% prcomp
plot_data = cbind(data.frame('PC1'=pca$x[,1], 'PC2'=pca$x[,2]), DE_info[DE_info$significant==TRUE,])
pos_zero = -min(plot_data$log2FoldChange)/(max(plot_data$log2FoldChange)-min(plot_data$log2FoldChange))
p = plot_data %>% ggplot(aes(PC1, PC2, color=log2FoldChange)) + geom_point(alpha=0.5) +
scale_color_gradientn(colours=c('#F8766D','#faa49e','white','#00BFC4','#009499'),
values=c(0, pos_zero-0.05, pos_zero, pos_zero+0.05, 1)) +
theme_minimal() + ggtitle('
PCA of differentially expressed genes') + # This is on purpose, PDF doesn't save well without this white space (?)
xlab(paste0('PC1 (',round(100*summary(pca)$importance[2,1],1),'%)')) +
ylab(paste0('PC2 (',round(100*summary(pca)$importance[2,2],1),'%)')) + theme(legend.position = 'bottom')
ggExtra::ggMarginal(p, type='density', color='gray', fill='gray', size=10)
rm(pos_zero, p)Separating the genes into these two groups: Salmon: under-expressed, aqua: over-expressed
plot_data = plot_data %>% mutate('Group'=ifelse(log2FoldChange>0,'overexpressed','underexpressed')) %>%
mutate('Group' = factor(Group, levels=c('underexpressed','overexpressed')))
rm(pca)List of DE genes
# Get genes names
getinfo = c('ensembl_gene_id','external_gene_id')
mart = useMart(biomart='ENSEMBL_MART_ENSEMBL', dataset='hsapiens_gene_ensembl', host='feb2014.archive.ensembl.org')
gene_names = getBM(attributes=getinfo, filters=c('ensembl_gene_id'), values=plot_data$ID, mart=mart) %>%
rename(external_gene_id = 'gene_name', ensembl_gene_id = 'ID')## Cache foundplot_data = plot_data %>% left_join(gene_names, by='ID')
kable(plot_data %>% dplyr::select(ID, gene_name, log2FoldChange, padj, Neuronal, Group) %>%
arrange(padj))| ID | gene_name | log2FoldChange | padj | Neuronal | Group |
|---|---|---|---|---|---|
| ENSG00000118514 | ALDH8A1 | -0.9022013 | 0.0067304 | 0 | underexpressed |
| ENSG00000188130 | MAPK12 | -0.5472590 | 0.0067304 | 0 | underexpressed |
| ENSG00000235750 | KIAA0040 | -1.0163837 | 0.0067304 | 0 | underexpressed |
| ENSG00000010404 | IDS | -0.4721840 | 0.0093103 | 0 | underexpressed |
| ENSG00000123358 | NR4A1 | 0.8078506 | 0.0356323 | 0 | overexpressed |
| ENSG00000134285 | FKBP11 | -0.4633866 | 0.0356323 | 0 | underexpressed |
| ENSG00000144406 | UNC80 | -0.2938665 | 0.0356323 | 0 | underexpressed |
| ENSG00000156711 | MAPK13 | -0.5719429 | 0.0356323 | 0 | underexpressed |
| ENSG00000162873 | KLHDC8A | -0.8858333 | 0.0402482 | 0 | underexpressed |
| ENSG00000090612 | ZNF268 | 0.3896882 | 0.0455581 | 0 | overexpressed |
| ENSG00000099814 | CEP170B | -0.2748351 | 0.0455581 | 0 | underexpressed |
| ENSG00000099917 | MED15 | -0.2858409 | 0.0455581 | 0 | underexpressed |
| ENSG00000110975 | SYT10 | -1.0485428 | 0.0455581 | 0 | underexpressed |
| ENSG00000136869 | TLR4 | 0.5320354 | 0.0455581 | 0 | overexpressed |
| ENSG00000138944 | KIAA1644 | -0.4311281 | 0.0455581 | 0 | underexpressed |
| ENSG00000139318 | DUSP6 | 0.7317736 | 0.0455581 | 0 | overexpressed |
| ENSG00000139767 | SRRM4 | -0.3288048 | 0.0455581 | 0 | underexpressed |
| ENSG00000154237 | LRRK1 | -0.4006294 | 0.0455581 | 0 | underexpressed |
| ENSG00000162944 | RFTN2 | 0.3844064 | 0.0455581 | 0 | overexpressed |
| ENSG00000178773 | CPNE7 | -0.4878353 | 0.0455581 | 0 | underexpressed |
| ENSG00000183741 | CBX6 | -0.2563460 | 0.0455581 | 0 | underexpressed |
| ENSG00000124507 | PACSIN1 | -0.2941277 | 0.0476343 | 1 | underexpressed |
| ENSG00000141441 | GAREM | 0.3352435 | 0.0476343 | 0 | overexpressed |
| ENSG00000147234 | FRMPD3 | -0.4979863 | 0.0476343 | 0 | underexpressed |
Effects of modifying the log fold change threshold
Not only we have very few DE genes, but we lose most of them from very low log Fold Change thresholds
#fc_list = seq(1, 1.04, 0.01)
fc_list = c(seq(1,1.01, 0.002), seq(1.01, 1.04, 0.01))
n_genes = nrow(datExpr)
# Calculate PCAs
datExpr_pca_samps = datExpr %>% data.frame %>% t %>% prcomp(scale.=TRUE)
datExpr_pca_genes = datExpr %>% data.frame %>% prcomp(scale.=TRUE)
# Initialice DF to save PCA outputs
pcas_samps = datExpr_pca_samps$x %>% data.frame %>% dplyr::select(PC1:PC2) %>%
mutate('ID'=colnames(datExpr), 'fc'=-1, PC1=scale(PC1), PC2=scale(PC2))
pcas_genes = datExpr_pca_genes$x %>% data.frame %>% dplyr::select(PC1:PC2) %>%
mutate('ID'=rownames(datExpr), 'fc'=-1, PC1=scale(PC1), PC2=scale(PC2))
pca_samps_old = pcas_samps
pca_genes_old = pcas_genes
for(fc in fc_list){
# Recalculate DE_info with the new threshold (p-values change) an filter DE genes
DE_genes = results(dds, lfcThreshold=log2(fc), altHypothesis='greaterAbs') %>% data.frame %>%
mutate('ID'=rownames(.)) %>% filter(padj<0.05)
datExpr_DE = datExpr %>% data.frame %>% filter(rownames(.) %in% DE_genes$ID)
n_genes = c(n_genes, nrow(DE_genes))
# Calculate PCAs
datExpr_pca_samps = datExpr_DE %>% t %>% prcomp(scale.=TRUE)
datExpr_pca_genes = datExpr_DE %>% prcomp(scale.=TRUE)
# Create new DF entries
pca_samps_new = datExpr_pca_samps$x %>% data.frame %>% dplyr::select(PC1:PC2) %>%
mutate('ID'=colnames(datExpr), 'fc'=fc, PC1=scale(PC1), PC2=scale(PC2))
pca_genes_new = datExpr_pca_genes$x %>% data.frame %>% dplyr::select(PC1:PC2) %>%
mutate('ID'=DE_genes$ID, 'fc'=fc, PC1=scale(PC1), PC2=scale(PC2))
# Change PC sign if necessary
if(cor(pca_samps_new$PC1, pca_samps_old$PC1)<0) pca_samps_new$PC1 = -pca_samps_new$PC1
if(cor(pca_samps_new$PC2, pca_samps_old$PC2)<0) pca_samps_new$PC2 = -pca_samps_new$PC2
if(cor(pca_genes_new[pca_genes_new$ID %in% pca_genes_old$ID,]$PC1,
pca_genes_old[pca_genes_old$ID %in% pca_genes_new$ID,]$PC1)<0){
pca_genes_new$PC1 = -pca_genes_new$PC1
}
if(cor(pca_genes_new[pca_genes_new$ID %in% pca_genes_old$ID,]$PC2,
pca_genes_old[pca_genes_old$ID %in% pca_genes_new$ID,]$PC2 )<0){
pca_genes_new$PC2 = -pca_genes_new$PC2
}
pca_samps_old = pca_samps_new
pca_genes_old = pca_genes_new
# Update DFs
pcas_samps = rbind(pcas_samps, pca_samps_new)
pcas_genes = rbind(pcas_genes, pca_genes_new)
}
# Add Diagnosis/SFARI score information
pcas_samps = pcas_samps %>% mutate('ID'=substring(ID,2)) %>%
left_join(datMeta, by=c('ID'='Dissected_Sample_ID')) %>%
dplyr::select(ID, PC1, PC2, fc, Diagnosis, Brain_lobe)
# pcas_genes = pcas_genes %>% left_join(SFARI_genes, by='ID') %>%
# mutate('score'=as.factor(`gene-score`)) %>%
# dplyr::select(ID, PC1, PC2, lfc, score)
# Plot change of number of genes
ggplotly(data.frame('lfc'=log2(fc_list), 'n_genes'=n_genes[-1]) %>% ggplot(aes(x=lfc, y=n_genes)) +
geom_point() + geom_line() + theme_minimal() + xlab('|Log Fold Change|') + ylab('Remaining Genes') +
ggtitle('Remaining genes when modifying filtering threshold'))rm(fc_list, n_genes, fc, pca_samps_new, pca_genes_new, pca_samps_old, pca_genes_old,
datExpr_pca_samps, datExpr_pca_genes)Samples
Note: PC values get smaller as Log2 fold change increases, so on each iteration the values were scaled so it would be easier to compare between frames
Coloured by Diagnosis:
LFC = -1 represents the whole set of genes, without any filtering by differential expression
Filtering out genes that are not differentially expressed (\(H_0:lfc=0\)) separates the two diagnosis groups perfectly using the first Principal Component
Increasing the LFC threshold doesn’t seem to improve the separation between groups much
ggplotly(pcas_samps %>% mutate(abs_lfc=ifelse(fc==-1,-1,round(log2(abs(fc)),3))) %>%
ggplot(aes(PC1, PC2, color=Diagnosis)) + geom_point(aes(frame=abs_lfc, ids=ID)) +
theme_minimal() + ggtitle('Samples PCA plot modifying filtering threshold'))Genes
if(!'fcSign' %in% colnames(pcas_genes)){
pcas_genes = pcas_genes %>% left_join(DE_info, by='ID') %>% mutate(fcSign = ifelse(log2FoldChange>0,'Positive','Negative'))
}
ggplotly(pcas_genes %>% mutate(abs_lfc=ifelse(fc==-1,-1,round(log2(fc),3))) %>%
ggplot(aes(PC1, PC2, color=fcSign)) + geom_point(aes(frame=abs_lfc, ids=ID, alpha=0.1)) +
theme_minimal() + ggtitle('Genes PCA plot modifying |LFC| filtering threshold'))